| Fire is a kind of natural disaster with high incidence,which occurs frequently in the city,forest area and wilderness,and has a great impact on the life and property safety of residents.Due to the suddenness and rapid spread of fire,early detection and warning are the key to reduce the damage caused by fire.With the continuous development of intelligent equipment,video image recognition technology has become one of the hot research directions,this technology can intelligently identify the location and scope of fire,for the relevant departments to create conditions for early rescue.Moreover,video image recognition technology has a wide range of application value,not only can be applied in urban buildings,public places and other scenes,but also can play a role in the forest,wilderness and other natural environment.At present,most of the video fire identification technologies at home and abroad aim at flame identification,but flame is not an important feature in the early stage of fire,and the effect in field application is not very ideal.Because in the early stage of fire,smoke is its most prominent feature,so it is more timely and fast to identify the fire by smoke.The main work of this paper includes the following aspects:(1)Fire smoke identification method based on traditional manual characteristics.Firstly,the image is converted to HSV color space to segment the suspected smoke area,and the foreground of the image is segmtioned to highlight the smoke features and reduce the interference items in the background.Then the LBP texture feature and HOG edge gradient feature of the suspected smoke area were extracted respectively for analysis,and the linear combination of them into a new feature vector.Finally,the obtained feature vector is input into the support vector machine to complete the fire smoke recognition.The experimental results show that the traditional manual feature fusion based fire smoke identification method has a better recognition effect on fire smoke.(2)Light smoke identification method based on depth feature fusion.The traditional manual feature fusion method can describe the characteristics of fire smoke well,but the feature extraction effect of light smoke in the early stage of fire is not very good because of the environmental interference.In order to solve this problem,data enhancement is performed on the fire smoke data set to increase the number of samples in the data set.Then VGGNet convolutional neural network and ResNet convolutional neural network were used to extract the light smoke characteristics respectively.Finally,the feature vectors obtained from the fusion of traditional manual features were fused with the depth features extracted by VGGNet and ResNet respectively.The complementarity between the depth features and traditional manual features was utilized to further improve the performance of the smoke recognition model.The experimental results show that the fire smoke recognition method based on depth feature fusion has a good recognition effect on light smoke,and also has a high recognition accuracy in complex environment.(3)Classification optimization based on multi-feature fusion method.First of all,on the basis of analyzing the experimental data of fire smoke,the difficulties faced by the current fire smoke data set are summarized.Then,according to the characteristics of fire smoke identification task,the improvement measures of convolutional neural network classification method are proposed.Finally,the full connection layer of convolutional neural network is combined with support vector machine,and it is adapted to the features of fire smoke image to complete the recognition task.Through data set testing and verification,the optimized classification method can effectively reduce the false alarm rate in the fire smoke identification task and has good timeliness,which provides an algorithm basis for fast and accurate fire warning. |
